Diffusion Tensor MR Imaging of the Neurologically Intact Human Spinal Cord

Subjects

Thirteen neurologically intact subjects (6 men and 7 women) between the ages of 19 and 40 years old (median age, 25 years) participated in the study. The number of subjects was chosen on the basis of a power analysis¹⁸ of average fractional anisotropy (FA) values with use of β = 0.9, α = 0.05, SD in FA of 0.1, and an effect size of 0.1 for comparisons. Subjects had no history of neurologic illness and were free from medical implants. We obtained informed consent before enrolling subjects. All procedures complied with the principles of the Declaration of Helsinki and were approved by the Institutional Review Boards of Marquette University and the Medical College of Wisconsin.

MR Imaging

We obtained axial DTI throughout the entire spinal cord (C1–L1) by using a standard single-shot, twice-refocused, spin-echo EPI pulse sequence.¹⁷ A CTL spine coil (GE Medical Systems, Milwaukee, Wis) and 1.5T clinical MR scanner (Signa Excite; GE Medical Systems) were used for all image acquisitions. We determined spinal levels using anatomic landmarks from sagittal fast low-angle shot (FLASH) locator MR images. DT images were acquired nonsequentially with TR, 6000 ms; TE, 96.3 ms; matrix size, 128 × 128; NEX, 1; FOV, 200 mm; section thickness, 5 mm; and no intersection gap (ie, contiguous). After reconstruction, images were zero-padded and interpolated to 256 × 256. Diffusion-weighted images (DWIs) were collected with b = 1500 s/mm², in 25 equidistant directions, and a single T2-weighted image (b = 0) was collected for each axial section.

After imaging, we calculated diffusion parameters and SNR from axial DT images of the spinal cord to identify normative diffusion values. Initially, the Analysis of Functional Neuroimages software (AFNI; available at http://afni.nimh.nih.gov/) was used to perform a Fourier transform-based affine registration of the 25 DWIs to the T2-weighted reference image to correct for bulk motion, eddy current, and susceptibility-related image distortions. AFNI was also used to calculate the 3 × 3 diffusion tensor from the 25 DWIs, and the eigenvalues of the tensor (λ₁, λ₂, λ₃) were identified for each voxel. The mean diffusivity (MD, or average of all 3 eigenvalues) was calculated to characterize the overall diffusion, and fractional anisotropy (FA) and volume ratio (VR)¹⁹ were calculated off-line as measures of diffusion anisotropy for each voxel with use of in-house MATLAB (Mathworks, Natick, Mass) algorithms. In addition, the transverse apparent diffusion coefficient (tADC, or average of the smallest 2 eigenvalues, λ₂ and λ₃) and longitudinal apparent diffusion coefficient (lADC, or largest eigenvalue, λ₁) were calculated to compare with literature values (see below). A total of 108 to 126 axial sections were obtained throughout the entire spinal cord, resulting in a total scan time of approximately 45 ± 10 SD minutes, depending on the height of the subject. Individual subject scan times were determined from the start of the first FLASH locator image to the last DWI image. We calculated T2-weighted SNR for every section in every subject and then averaged across spinal levels for each subject by using the method outlined by Kaufman et al.²⁰ This resulted in 1 SNR distribution per subject per level.

Data Description for Regions of Interest

Manual segmentation of spinal cord regions of interest (ROIs) was performed for the whole cord and for separate identification of gray and white matter regions. First, we manually segmented the spinal cord from the CSF by using the T2-weighted images for a whole cord analysis (ie, no gray and white matter segmentation). ROIs for the whole cord analysis were placed within the spinal cord such that at least 2 voxels around the edge of the cord were excluded to assure no partial volume contamination from the surrounding CSF. Whole cord segmentation may benefit future studies conducted in imaging planes or with use of pulse sequences where gray and white matter differentiation is not possible or may be particularly difficult. For example, gray and white matter differentiation in images collected within the sagittal plane is not typically feasible. Thus, these studies will benefit from diffusion values related to the whole cord instead of diffusion characteristics related to individual ROIs within the cord. In addition, diffusion values and anisotropy indices from the CSF surrounding the cord were collected from each image section.

Manual identification of targeted ROIs within the spinal cord were selected from FA color images of the spinal cord to identify white matter (WM) and gray matter (GM) regions. For GM and WM ROIs, regions were selected within the spinal cord such that at least 2 voxels around the edge of the tissue of interest were excluded to assure no partial volume contamination. This method of segmentation may benefit future studies aimed at characterizing changes in diffusion characteristics within particular regions of the spinal cord. For example, investigators studying motor neuron plasticity after injury to the spinal cord may be interested in changes in GM diffusion properties that result from specific treatments or rehabilitation strategies. Although more targeted ROIs could have been created to represent individual WM spinal tracts, image resolution at 1.5T restricts these tracts to only a few voxels per image and thus are susceptible to partial volume contamination. Also, we did not control for the motion of the spinal cord; therefore, confident estimates of diffusion values within smaller spinal tracts were not feasible.

For the whole cord analysis, histograms of the diffusion characteristics for each spinal level were created to represent the results. Sections in the same transverse plane as intervertebral disks were excluded from analysis to decrease the effects of susceptibility-related distortions.²¹ The remaining sections were pooled to create a single distribution per spinal segment for each subject (approximately 3–5 sections per segment per subject, or a total of approximately 200–600 voxels per segment per subject). Finally, distributions for all subjects were pooled across respective spinal levels. The total number of voxels represented in each diffusion distribution for a particular spinal level reported in this study ranged from approximately 2500 to 5000 voxels.

Group statistics (mean and SD) were then calculated for the diffusion parameters. For each subject, the mean eigenvalues, FA, and VR were calculated for each segment and each ROI (whole cord, WM, GM, and CSF). Then, we calculated the group means and SDs by using the subject means. The data were represented by plotting the values against segmental level.

There is currently no criterion standard for spinal cord segmentation in MR imaging. In our study, we have chosen to use manually defined ROIs, which is the method of choice in most DTI studies of the spinal cord.^{3,8,10,12,14,22} Other methods of segmentation used in DTI analysis of the human spinal cord include thresholding the diffusion or anisotropy images²³ and fuzzy-logic–based tissue classification.^6,24 These methods seem to be sufficient; however, they may differ in their performance if diffusion characteristics change throughout the length of the cord.

To validate the manual segmentation technique, we calculated the percentage of agreement between manual segmentation and the 2 other segmentation techniques. We calculated the percentage agreement for each section by first manually cropping the 256 × 256 images to 128 × 128. Then, segmentation was implemented, and the number of similarly labeled voxels from the 2 techniques was found. The total percentage of agreement was calculated by dividing the number of similarly labeled voxels by the total number of voxels in the cropped image (128 × 128 image = 16,384 voxels). First, we compared manual segmentation of the spinal cord from the CSF with segmentation by using an FA threshold of 0.3, resulting in an average 78% agreement (median, 81%). Second, manual segmentation of the entire spinal cord from the CSF was compared with a fuzzy-logic–based segmentation described elsewhere.²⁴ This comparison resulted in an average of 81% agreement (median, 85%). To compare ROIs within the spinal cord (GM, WM), we compared manual segmentation with fuzzy-logic–based segmentation.²⁴ The percent agreement between these techniques for GM, WM, and CSF averaged 61%, 72%, and 91%, respectively. Because manual segmentation produced agreement in most scenarios, values obtained from manually selected ROIs were used in the statistical analysis.

Statistics

To determine if SNR varied across the length of the spinal cord, we used a one-way repeated-measures analysis of variance (ANOVA) to evaluate differences in SNR across spinal levels (fixed factor: spinal level; random factor: subject). In a similar fashion, each eigenvalue (λ₁, λ₂, λ₃), MD, FA, and VR were evaluated across the length of the spinal cord with use of a one-way repeated-measures ANOVA if the data were normally distributed. If the dataset failed the normality test, a Friedman repeated measures nonparametric ANOVA on ranks was performed. We conducted multiple post hoc comparisons between spinal levels for all measurements by using Tukey tests for multiple comparisons, either in nonparametric or parametric form depending on the ANOVA performed. We conducted post hoc comparisons by using Tukey tests. The level of significance used for all statistical tests was α = 0.05. Minitab 14 software (Minitab, State College, Pa) was used for the statistical analyses.

Comparison With Published Diffusion Characteristics

To address the fundamental question as to whether diffusion characteristics can be generalized across imaging platforms, pulse sequences, and postprocessing algorithms, we identified an estimate of MD across multiple published studies of spinal cord diffusion imaging (see Table for summary). Using data from 13 studies, we estimated signal intensity attenuation by using the diffusion equation: Here, (S/S₀) is signal intensity attenuation, b is the diffusion weighting in s/mm², and MD is mean diffusivity in mm²/s. Once an estimate of signal intensity attenuation was calculated from each study, data were plotted as a function of b-value and a regression was used to estimate MD across studies. Many of the diffusion values used in this calculation were averages across the spinal cord, and differences across segmental levels were not reported. Therefore, we did not distinguish between segments when performing the regression analysis, assuming MD is relatively consistent throughout the length of the spinal cord.

Comparison With Published Diffusion Characteristics

The estimated signal intensity attenuation was plotted versus diffusion weighting (b-value) for the published values (Fig 4), with results consistent with the expected exponential relationship, despite differences in imaging platform and pulse sequence. The subsequent logarithmic regression resulted in a diffusivity of 0.997 × 10⁻³ mm²/s with R² = 0.842 and P < .001. This value was slightly higher than the MD measured in our study (0.83 × 10⁻³ mm²/s) but within 1 SD of the mean. (The SD of MD in our study was 0.2 × 10⁻³ mm²/s, measured across subjects and across spinal levels.)

• Download figure
• Open in new tab
• Download powerpoint

Fig 4.

Regression analysis of mean apparent diffusion coefficients from DTI studies of the human spinal cord with use of a variety of pulse sequences. Numbers represent identification numbers and are located in the accompanying table. Results of logarithmic regression are also shown. Solid circle indicates average signal intensity attenuation for this study.

Limitations

Despite obtaining similar average diffusion measurements to those reported in the current literature, many potential limitations to our study remained. The use of manual ROI selection may have introduced partial volume contamination, primarily in regions where the spinal cord cross-sectional area was smallest and in the lower thoracic regions where spinal nerves may run along the spinal cord. Despite a high percentage agreement between manual ROI selection and other techniques (see Methods section), a need for more precise segmentation is necessary. The use of fuzzy-logic–based tissue classification has recently been proposed as a method for delineating GM and WM in the spinal cord^6,24; however, its usefulness in pathologic conditions still requires validation.⁷

Another limitation to our study was the lack of cardiac and respiratory gating during image acquisition. Cardiac and respiratory gating during DT image acquisition substantially reduces errors related to diffusion measurements^30,31 because the spinal cord moves in the craniocaudal direction synchronously with the cardiac cycle. This is of particular interest when sagittal images of the spinal cord are obtained. Although motion of the spinal cord is lower in the transverse plane, the lack of cardiac and respiratory gating in our study was a possible limitation.

Susceptibility-related image artifacts have limited the clinical use of DTI in the spinal cord. These artifacts arise from magnetic susceptibility differences between the spinal cord and the surrounding bony vertebral column. Although we did observe slight geometric distortions (Fig 1), we were careful not to use sections that may have been strongly affected. In particular, the regions of the spinal cord most sensitive to changes in B0 from susceptibility inhomogeneity are located proximal to the vertebral disk (because of the spinous process from the rostral vertebral level).²¹ Furthermore, we used a Fourier-transform–based image registration algorithm to correct geometric distortions between the T2-weighted and DWIs. Regardless, the lack of B0 correction for susceptibility artifacts for both T2-weighted and DWIs may have distorted the images of our study.

A final limitation to our study was the age range represented by this normative dataset. MD increases and FA decreases with increasing age, though the magnitude of the changes is minimal. Between ages 25 and 85 years, the average MD increases 0.1 × 10⁻³ mm²/s, and mean FA decreases approximately 0.04,³ well within the ranges represented in our dataset. Although the magnitude of changes in diffusion characteristics is expected to change only slightly during normal aging, discretion should be used when comparing our data with older populations.